On Fri, Jan 26, 2024 at 02:28:42PM +0100, Christoph Hellwig wrote: > One thing I don't really like yet is that xfile is still based on > folios and not pages. The main stumbling block for that is the > mess around the hwpoison flag - that one still is per-file and not > per-folio, and shmem checks it weirdly often and not really in hwpoison is per page not per file. That's intrinsic to, well, hardware poison, it affects an entire page (I'd love to support sub-page poison, but not enough to spend my time working on memory-poison.c). In general, I think there's a lack of understanding of hwpoison, and I include myself in that. Mostly I blame Intel for this; limiting the hardware support to the higher end machines means that most of us just don't care about it. Why even bother checking for hwpoison in xfiles? If you have flaky hardware, well, maybe there's a reason you're having to fsck, and crashing during a fsck might encourage the user to replace their hardware with stuff that works.
